Vibe coding i praktiken – utveckling avett Python‑baserat verktyg förautomatiserad modellering ochdimensionering av stålplattformar

Sammanfattning

Automation and digitalization are becoming increasingly important in structural engineering,and the use of artificial intelligence is pushing the limits of what can be achieved. This paperinvestigates how the modelling, dimensioning and optimization of steel platforms in an earlydesign stage can be automated using a Python-script, developed with the method called vibecoding.The code editor Cursor was used due to its integration of AI in the development process.Structural dimensioning of beams and columns was performed using a finite element method(FEM) program. To create a structural model in Tekla Structures, an IFC file of the steelplatform with optimized cross-sections was generated directly by the Python script.The quantitative results show good agreement between the results from the Python script andhand calculations, with only minor differences, such as a maximum deviation of 0.8 mm indeflection (SLS) and 4.9 percentage points in utilisation ratio (ULS), which are considerednegligible. This indicates that it is possible to develop a reliable Python script usingthe development method vibe coding, if user performs the required verification procedure.The qualitative results, obtained through a semi-structured focus group interview with structuralengineers, show that the Python script is perceived as a useful tool in the early design stages,with the potential to reduce material usage and shorten the early design phase. At the same time,the script was regarded as a proof of concept, and several limitations were identified, primarilyrelated to geometric flexibility, the handling of more complex load cases, and a lack of clarityregarding how to execute the Python-script.The overall results demonstrate that the vibe coding approach is a viable method for developingPython-script for this application. However, several issues were identified, more complex codechanges required highly specific prompts and the use of incorrect technical terminology led tofailure to produce the wanted changes.